rescale-ai-client 0.3.4

Standalone HTTP client for hosted and local Rescale AI inference servers

rescale-ai-client

Standalone Python client for Rescale AI hosted and local inference servers. Communicates over the shared HTTP API. No rescale_ai package dependency.

Installation

pip install rescale-ai-client

For mesh loading, prediction plotting, or VTK mesh helpers (.vtp and .vtm):

pip install "rescale-ai-client[mesh]"

For interactive 3D notebook plots:

pip install "rescale-ai-client[mesh,jupyter]"

This extra also installs jupyter_bokeh, which Panel needs for interactive rendering in VSCode notebooks.

For development (do not run unless making local changes to the rescale-ai-client library)

pip install -e ".[dev]"

License Request

uvx rescale-ai-client request-license --email [e-mail]

Writes license-request.json for the current user and machine.

License Registration

uvx rescale-ai-client register

Registers the issued license for the current user. Without a path, the command looks for a valid license in ~/Downloads, then the current directory, then prompts for the absolute path to the license file.

You can also pass the license path explicitly:

uvx rescale-ai-client register <path_to_license_file>

Quickstart

Launch Local Jupyter Notebook:

cd <path_to_workspace>/rescale-ai-client
source .venv/bin/activate
python -m ensurepip --upgrade
python -m pip install --upgrade pip
pip install uv
jupyter notebook

See example/notebooks/quickstart.ipynb for an example Jupyter notebook to get started. The directory contains other more advanced examples.

Retrieve Server-Available Models:

from rescale_ai_client import InferenceClient

client = InferenceClient.connect(base_url="http://127.0.0.1:65432") # Connects to local Inference Runner. Modify to connect to remote Host & Port
catalog = client.list_models()

model = catalog.model("MyModel").latest()
prediction = client.run_inference(
    "/path/to/case.vtm",
    parameters={"mach": 0.82},
    model_name=model.model_name,
    version_number=model.version_number,
)

print(prediction.global_predictions)
prediction.save("/tmp/result.vtm")

No-mesh Core Install, Summary-Only Inference:

from rescale_ai_client import InferenceClient

client = InferenceClient.connect(base_url="http://127.0.0.1:65432") # Connects to local Inference Runner. Modify to connect to remote Host & Port
summary = client.run_inference_summary(
    "/path/to/case.vtm",
    parameters={"mach": 0.82},
    model_name="MyModel",
    version_number=7,
)

print(summary.global_predictions)
print(summary.prediction_id)

Offline Install Bundle:

This installation path used when downloading an inference bundle directly from the Web UI.

from rescale_ai_client import InferenceClient

client = InferenceClient.connect(base_url="http://127.0.0.1:65432") # Connects to local Inference Runner. Modify to connect to remote Host & Port

installed = client.install_bundle(
    "/path/to/MyModel_v0_local_inference.zip"
)

prediction = client.run_inference(
    "/path/to/case.vtm",
    parameters={"mach": 0.82},
    model_version_path=installed.model_version_path,
)

print(installed.model_name, installed.version_number)
print(prediction.global_predictions)

An ensemble bundle loads as LoadedEnsemble and is selected by its immutable ensemble identity. Members execute sequentially, and uncertainty arrays can be requested alongside the ensemble mean:

from rescale_ai_client import EnsembleLocator, InferenceClient

client = InferenceClient.connect()
installed = client.install_bundle("/path/to/ensemble.zip")

prediction = client.run_inference(
    "/path/to/case.vtm",
    model=EnsembleLocator(ensemble_id=installed.ensemble.ensemble_id),
    include_standard_deviation=True,
    ensemble_node_output="mean",
)

for name, distribution in prediction.ensemble_global_predictions.items():
    print(name, distribution.mean, distribution.standard_deviation)
    print(distribution.member_predictions)

Global ensemble outputs always include their mean, standard deviation, and member-value vector. Node arrays default to the ensemble's primary member; set ensemble_node_output="mean" to generate mean node predictions. Node standard-deviation arrays are only generated when mean node output and include_standard_deviation=True are both requested.

extensions for Optional Future Data

Most JSON request methods accept an optional extensions={...} payload. This is the additive escape hatch for future server features that should not force a client upgrade.

Example: Send Optional Request Hints During Prediction:

from rescale_ai_client import InferenceClient

client = InferenceClient.connect(base_url="http://127.0.0.1:65432") # Connects to local Inference Runner. Modify to connect to remote Host & Port

summary = client.run_inference_summary(
    "/path/to/case.vtm",
    parameters={"mach": 0.82},
    model_name="MyModel",
    version_number=7,
    extensions={
        "rescale": {
            "ui_session": "demo-123",
            "include_debug": True,
        }
    },
)

print(summary.extensions)
print(summary.warnings)

Example: Read Capability-Style or Server-Added Metadata from Typed Responses:

from rescale_ai_client import InferenceClient

client = InferenceClient.connect(base_url="http://127.0.0.1:65432")

server_info = client.server_info()
print(server_info.capabilities)

metadata = client.get_model_version_metadata(
    model_name="MyModel",
    version_number=7,
)
print(metadata.extensions)
print(metadata.warnings)

Run a Saved System

Saved systems created in the Design Studio are available through the local inference server. Inputs may be partial; omitted values use the defaults stored in the system definition. Passing the catalog hash prevents accidentally running a definition that changed after it was inspected:

from rescale_ai_client import InferenceClient

client = InferenceClient.connect()
catalog = client.list_systems()
system = catalog.system("booth")

result = client.run_system(
    system.system_id,
    {"x1": 1.0, "x2": 3.0},
    expected_system_hash=system.system_hash,
)

print(result.outputs)
print(result.resolved_models)

get_system(system_id) returns the declared inputs, outputs, components, and validation state. A run records the exact model versions resolved at its start, including models configured as latest. System catalog and execution endpoints are intentionally available only through a loopback connection (direct local access or an SSH tunnel).

Design rules:

• extensions are optional and additive
• unknown request extensions are ignored by the server
• unknown response extensions are ignored by clients
• if a field becomes core contract, it should graduate into a typed top-level field

Optimization

The rescale_ai_client.optimization package turns mesh inference into a design-optimization loop. Define an Objective (built-ins cover global scalar outputs, point-field reductions, and multi-model chains) and hand it to a search driver. Every driver minimizes ObjectiveResult.value (use maximize=True on the built-in objectives to flip the sign) and returns a result carrying the full evaluation history, which the plot_* helpers can render.

Install the extra:

pip install "rescale-ai-client[optimization]"   # adds SciPy
pip install "rescale-ai-client[cma]"             # also adds CMA-ES support

Available drivers:

Driver	Kind	Use when
minimize_scipy	Local (L-BFGS-B / SLSQP / TNC)	A good starting point is known and the response is smooth
sqp_minimize	Local SQP (SLSQP / trust-constr) with nonlinear constraints	The problem has constraints (e.g. max_stress <= 500); warm-start from a global driver
differential_evolution	Global, population	Multi-modal landscape, evaluations are cheap
random_search	Baseline sampling	A quick lower bound / sanity check
bayesian_optimization	Global, GP surrogate + Expected Improvement	Each evaluation is an expensive inference call (most sample-efficient)
cma_es	Global evolution strategy (needs [cma])	Noisy or ill-conditioned, multi-modal objectives

from rescale_ai_client import InferenceClient
from rescale_ai_client.optimization import (
    GlobalOutputObjective,
    bayesian_optimization,
    cma_es,
    minimize_scipy,
    random_search,
)

client = InferenceClient.connect(base_url="http://127.0.0.1:65432")
obj = GlobalOutputObjective(
    client,
    model_name="MyModel",
    version=0,
    vtp="/path/to/case.vtm",
    output_name="Cd",
    maximize=False,
)
bounds = {"AngleBeta1": (10.0, 35.0), "RPM": (25000.0, 32000.0)}

# Sample-efficient global search — best for expensive inference objectives.
result = bayesian_optimization(obj, bounds=bounds, n_calls=25, seed=0, verbose=True)
print(result.x, result.fun)

# Other drivers share the same call/return shape:
# minimize_scipy(obj, x0={...}, bounds=bounds)
# cma_es(obj, x0={...}, bounds=bounds)            # requires the [cma] extra
# random_search(obj, bounds=bounds, n_calls=50)

Constrained optimization with SQP

sqp_minimize adds nonlinear constraints. Each Constraint receives both the parameters and the cached ObjectiveResult at that point, so a constraint on a model output reuses the objective's inference (no extra call):

from rescale_ai_client.optimization import Constraint, sqp_minimize

constraints = [
    # Keep a predicted global output under a limit (read from the cached result).
    Constraint(
        fn=lambda params, result: result.info["global_outputs"]["max_stress"],
        upper=500.0,
        name="max_stress",
    ),
    # A purely geometric/parameter constraint ignores the result argument.
    Constraint(fn=lambda params, result: params["t1"] - params["t2"], lower=0.0),
]

result = sqp_minimize(
    obj,
    x0={"AngleBeta1": 20.0, "RPM": 29607.0},
    bounds=bounds,
    constraints=constraints,
    method="SLSQP",          # or "trust-constr" for harder problems
    verbose=True,
)
print(result.x, result.fun, result.feasible, result.constraint_values)

Because SQP estimates gradients by finite differences, it's most effective on a smooth response and warm-started from a global driver's best point.

API

InferenceClient.connect(*, base_url=None, host=None, port=None, url=None, timeout=600.0)

Factory method. If omitted, defaults to local loopback at http://127.0.0.1:65432. base_url first. host / port and url remain convenience inputs. Scheme-less non-local endpoints default to HTTPS; explicit HTTP for non-local endpoints raises an InsecureConnectionWarning and stores the message on client.insecure_transport_warning.

// Give example here: connecting to my most recent Workstation

InferenceClient.connect_via_ssh(*, ssh_user_host, remote_port=65432, ...)

SSH tunnel helper. ssh_user_host required. If remote_port is omitted, defaults to 65432.

run_inference(vtp, parameters=None, *, model=None, model_version_path=None, model_name=None, model_uuid=None, version_number=None, ...)

Canonical Python entrypoint. model= accepts either ModelVersionLocator or EnsembleLocator. The vtp argument name is kept for compatibility, but path inputs may be .vtp or .vtm; .vtm paths are sent to the inference server so block structure is preserved. Same shape for hosted and local servers. CamelCase aliases still work, but new code should use snake_case. JSON request methods also accept extensions=... for optional future payloads.

run_inference_summary(...)

Lightweight entrypoint for core installs without the optional mesh dependencies. Returns typed prediction metadata plus global scalar outputs without constructing a pyvista mesh.

list_models(...)

Returns a typed ListModelsResult with helpers like .model("name"), .get("name", version_number=7), and .model("name").latest().

list_systems() / get_system(system_id)

Lists saved Design Studio systems or returns one typed SystemDescription.

run_system(system_id, inputs=None, *, expected_system_hash=None, require_convergence=False, extensions=None)

Runs a saved system on the local inference server and returns a typed SystemRunResult containing scalar outputs, resolved model versions, iteration diagnostics, warnings, and reproducibility hashes.

inspect_bundle(...) / install_bundle(...)

Inspect, upload, and install an offline local-inference bundle zip into the local installer. install_bundle(...) is idempotent for identical content. Ensemble IDs are immutable, so installing different content under an existing ensemble ID is rejected.

get_model_version_metadata(version_number, *, model_name=None, model_uuid=None, ...)

Fetches typed metadata for one model version.

get_loaded_model()

Returns the server's current active-model bookkeeping status.

server_info()

Returns typed inference server capability and compatibility metadata:

• api_version
• supported_api_versions
• schema_revision
• server_build
• capabilities

Client

Compatibility alias for one release. Use InferenceClient for new code.

Advanced

Protocol Sync

The committed wire DTOs under src/rescale_ai_client/_generated/ are generated from the closed-source inference server OpenAPI schema snapshot at openapi/inference-server.openapi.json.

Update flow:

# from the closed-source rescale_ai repo
PYTHONPATH=/path/to/rescale_ai \
  /path/to/rescale_ai/.venv/bin/python \
  - <<'PY'
from pathlib import Path
from rescale_ai.inference.server.openapi_compat import export_openapi_schema

export_openapi_schema(
    Path("/path/to/rescale-ai-client/openapi/inference-server.openapi.json")
)
PY

# from this repo
python scripts/generate_protocol_models.py
python scripts/generate_protocol_models.py --check

Releasing

Releases are published by GitHub Actions with uv build and uv publish.

One-Time Setup:

1. In GitHub, create an environment named pypi under Settings -> Environments.
2. In PyPI, add a trusted publisher for:
   • Owner: rescale
   • Repository: rescale-ai-client
   • Workflow: release.yml
   • Environment: pypi

No PyPI API token needs to be stored in GitHub secrets for this workflow. PyPI issues a short-lived publishing credential to the tagged GitHub Actions run through OIDC.

Publish a Release:

uv version 0.1.1
git commit -am "Release 0.1.1"
git tag -a v0.1.1 -m v0.1.1
git push origin main
git push origin v0.1.1

The workflow runs tests, verifies generated protocol models, builds both wheel and source distributions, smoke-tests both artifacts, and publishes to PyPI.

License

Apache-2.0